Evaluation and iteration

An unevaluated agent is an unreliable agent. The evaluation must cover the full pipeline: retrieval quality, extraction accuracy, citation correctness, synthesis coherence, and guardrail effectiveness. Each dimension needs its own metric and test set.